足球视频分析系统

作者

• 陆徐东 [email protected]
• 方天宬 [email protected]

简介

本项目是SJTU 21-22学年CS386 数字图像处理课程的大作业，本文是足球视频分析系统的参考文档。我们主要实现了以下功能：

1. 基于Yolo v5和PastaNet搭建了足球视频的分析神经网络，能够对球员位置、球员姿态和动作进行识别，也能对球队战术进行初步识别
2. 基于Qt6搭建了一套足球分析系统，包括服务端和客户端：客户端上传视频到服务端，分析完成后再下载结果并展示

使用方法

• 服务端：

  1. 需要一台装有NVIDIA20系列显卡，并且装有cuda10.2的Linux电脑（如果你打算用CPU运行神经网络，没有显卡也可以）
  2. 配置python环境，输入conda env create -n activity2vec -f DIP/HAKE-Action-Torch-Activity2Vec/activity2vec.yaml和conda env create -n yolo -f DIP/Yolov5_DeepSort_Pytorch/yolo.yaml
  3. 在Linux环境下用Qt6编译src/server-console/server-console.pro，如果是在docker中，那么还需用ldd命令找到所需的库文件，将编译好的可执行文件和库文件一起拷贝到docker
  4. 修改DIP文件夹下面三个.sh脚本，将其中的$PYTHON_PATH改成自己conda环境中对应的python位置
  5. 将编译好的server-console放到DIP文件夹下，运行之

• 客户端：

  1. 下载并安装Qt6
  2. 用Qt6打开src/layouts/basiclayouts.pro，编译之
  3. 把DIP/GUI中的get_place.py打包成get_place.exe，并与第二步编译好的文件放在同一目录下
  4. 运行第二步编译好的文件

文件夹

DIP  神经网络方面
|-- GUI            人工校准GUI
|-- inputfile      输入文件
|-- Yolov5_DeepSort_Pytorch
|-- HAKE-Action-Torch-Activity2Vec
...

src  图像界面方面
|-- layouts        客户端
|-- server-console 服务端

report 报告

注意

• 我们在此处没有提供全套DIP文件夹，它足足有7.2G，您可以根据下面的链接下载环境

  链接: https://pan.baidu.com/s/1PiAyDIr59o5IvgcjAnUylw  密码: 0gso
  --来自百度网盘超级会员V5的分享
  

Football Video Analyse System

• Xudong Lu [email protected]
• Tiancheng Fang [email protected]

Introduction

This project is a major assignment of cs386 digital image processing course of SJTU 21-22 academic year. This tutorial is a reference document for football video analysis system. We mainly realize the following functions:

1. We build a football video analysis neural network, which can identify the player's position, player's posture and action, and also preliminarily identify the team's tactics.
2. We construct a football analysis system based on Qt-6, including server and client: the client uploads the video to the server, downloads the results and displays them after analysis.

Usage

• Server:

  1. You need a CUDA 10.2 Linux computer with NVIDIA 20 series graphics card (If you plan to run neural networks with CPU, you can do it without a graphics card)
  2. Build the python environment, enter conda env create -n activity2vec -f DIP/HAKE-Action-Torch-Activity2Vec/activity2vec.yaml and conda env create -n yolo -f DIP/Yolov5_DeepSort_Pytorch/yolo.yaml
  3. Compile src/server-console/server-console.pro in Linux Qt6. If you decide to run it in docker, you also need ldd command to find the required library, then copy the executable file and the library to docker
  4. Modify the three .sh script in folder DIP, change $PYTHON_PATH to the corresponding Python location in your conda environment
  5. Put the executable server-console into the DIP folder, then run it

• Client:

  1. Download and install Qt6
  2. Open src/layouts/basiclayouts.pro with Qt6, then compile it
  3. Pack the DIP/GUI/get_place.py to get_place.exe and put it in the same directory as the files compiled in step 2
  4. Run the file compiled in step 2

Directory

DIP  # about neural network
|-- GUI            # calibrate GUI
|-- inputfile      
|-- Yolov5_DeepSort_Pytorch
|-- HAKE-Action-Torch-Activity2Vec
...

src  # about GUI
|-- layouts        # client
|-- server-console # server

report 报告

Note

• We don't provide a full set of DIP folders here. It takes up 7.2G of space. You can download the environment according to the link below:

  URL: https://pan.baidu.com/s/1PiAyDIr59o5IvgcjAnUylw
  password: 0gso